def test_sgd_regressor_rk_as(loss):
rng = np.random.RandomState(0)
transform = RandomKernel(n_components=100,
random_state=0,
kernel='all_subsets')
X_trans = transform.fit_transform(X)
y, coef = generate_target(X_trans, rng, -0.1, 0.1)
y_train = y[:n_train]
y_test = y[n_train:]
_test_regressor(transform, y_train, y_test, X_trans, loss=loss)
def test_anova_kernel(dist, degree, kernel):
# compute exact kernel
gram = anova(X, Y, degree)
# approximate kernel mapping
rk_transform = RandomKernel(n_components=1000, random_state=0,
kernel=kernel, degree=degree,
distribution=dist, p_sparse=0.5)
X_trans = rk_transform.fit_transform(X)
Y_trans = rk_transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
error = gram - kernel_approx
assert np.abs(np.mean(error)) < 0.0001
assert np.max(error) < 0.001 # nothing too far off
assert np.mean(error) < 0.0005 # mean is fairly close
# sparse input
X_trans_sp = rk_transform.transform(X_sp)
assert_allclose_dense_sparse(X_trans, X_trans_sp)
# sparse output
if dist == "sparse_rademacher":
rk_transform.dense_output = False
X_trans_sp = rk_transform.transform(X_sp)
assert issparse(X_trans_sp)
assert_allclose_dense_sparse(X_trans, X_trans_sp.toarray())
else:
rk_transform.dense_output = False
X_trans_sp = rk_transform.transform(X_sp)
assert not issparse(X_trans_sp)
assert_allclose_dense_sparse(X_trans, X_trans_sp)
def test_regressor_rk_as(normalize, loss):
rng = np.random.RandomState(0)
# approximate kernel mapping
transformer = RandomKernel(n_components=100, random_state=0,
kernel='all_subsets')
X_trans = transformer.fit_transform(X)
y, coef = generate_target(X_trans, rng, -0.1, 0.1)
y_train = y[:n_train]
y_test = y[n_train:]
_test_regressor(transformer, X_train, y_train, X_test, y_test, X_trans,
normalize=normalize, loss=loss)
def test_sgd_regressor_rk(loss):
rng = np.random.RandomState(0)
for degree in range(2, 5):
transform = RandomKernel(n_components=100,
random_state=0,
degree=degree)
X_trans = transform.fit_transform(X)
y, coef = generate_target(X_trans, rng, -0.1, 0.1)
y_train = y[:n_train]
y_test = y[n_train:]
_test_regressor(transform, y_train, y_test, X_trans, loss=loss)
def test_sparse_rademacher(p_sparse):
# approximate kernel mapping
for p_sparse in [0.9, 0.8, 0.7, 0.6, 0.5]:
rk_transform = RandomKernel(n_components=1000, random_state=0,
kernel='anova',
distribution='sparse_rademacher',
p_sparse=p_sparse)
X_trans = rk_transform.fit_transform(X)
nnz_actual = rk_transform.random_weights_.nnz
nnz_expected = X.shape[1]*rk_transform.n_components*(1-p_sparse)
assert_almost_equal(np.abs(1-nnz_actual/nnz_expected), 0, 0.1)
def test_classifier_rk(normalize, loss, degree):
rng = np.random.RandomState(0)
# approximate kernel mapping
transformer = RandomKernel(n_components=100, random_state=0,
degree=degree)
X_trans = transformer.fit_transform(X)
y, coef = generate_target(X_trans, rng, -0.1, 0.1)
y_train = y[:n_train]
y_test = y[n_train:]
_test_classifier(transformer, X_train, np.sign(y_train), X_test,
np.sign(y_test), X_trans, normalize=normalize,
loss=loss)
def test_compact_random_feature_random_kernel_all_subsets(down):
# approximate kernel mapping
transform_up = RandomKernel(kernel='all_subsets',
random_state=0,
n_components=100)
transform_down = down(n_components=50, random_state=0)
X_trans_naive = transform_down.fit_transform(transform_up.fit_transform(X))
transform_up = RandomKernel(kernel='all_subsets',
random_state=0,
n_components=100)
transform_down = down(n_components=50, random_state=0)
transformer = CompactRandomFeature(transformer_up=transform_up,
transformer_down=transform_down)
X_trans = transformer.fit_transform(X)
assert_allclose(X_trans_naive, X_trans)
def test_sgd_classifier_rk(loss):
rng = np.random.RandomState(0)
for degree in range(2, 5):
transform = RandomKernel(n_components=100,
random_state=0,
degree=degree)
X_trans = transform.fit_transform(X)
y, coef = generate_target(X_trans, rng, -0.1, 0.1)
y_train = y[:n_train]
y_test = y[n_train:]
_test_classifier(transform,
np.sign(y_train),
np.sign(y_test),
X_trans,
max_iter=100,
eta0=.1,
loss=loss)
def test_all_subsets_kernel(dist):
# compute exact kernel
p_sparse = 0.5
kernel = all_subsets(X, Y)
# approximate kernel mapping
rk_transform = RandomKernel(n_components=5000, random_state=0,
kernel='all_subsets',
distribution=dist, p_sparse=p_sparse)
X_trans = rk_transform.fit_transform(X)
Y_trans = rk_transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
error = kernel - kernel_approx
if dist == 'sparse_rademacher':
nnz = rk_transform.random_weights_.nnz
nnz_expect = np.prod(rk_transform.random_weights_.shape)*p_sparse
nnz_var = np.sqrt(nnz_expect * (1-p_sparse))
assert np.abs(nnz-nnz_expect) < 3*nnz_var
assert np.abs(np.mean(error)) < 0.01
assert np.max(error) < 0.1 # nothing too far off
assert np.mean(error) < 0.05 # mean is fairly close
X_trans_sp = rk_transform.transform(X_sp)
assert_allclose_dense_sparse(X_trans, X_trans_sp)
if lkrf.remove_bases():
X_trans_removed = lkrf.transform(X)
assert_almost_equal(X_trans_removed.shape[1], n_nz)
indices = np.nonzero(lkrf.importance_weights_)[0]
assert_almost_equal(X_trans_removed, X_trans[:, indices])
params = [
RBFSampler(n_components=128, random_state=0),
RandomFourier(n_components=128, random_state=0),
RandomFourier(n_components=128, random_state=0, use_offset=True),
OrthogonalRandomFeature(n_components=128, random_state=0),
OrthogonalRandomFeature(n_components=128, random_state=0,
use_offset=True),
RandomMaclaurin(random_state=0),
RandomKernel(random_state=0)
]
@pytest.mark.parametrize("transformer", params)
def test_lkrf_chi2(transformer, rho=1):
_test_learning_kernel_with_random_feature('chi2', transformer, rho)
def test_lkrf_chi2_origin():
_test_learning_kernel_with_random_feature('chi2_origin')
@pytest.mark.parametrize("transformer", params)
def test_lkrf_kl(transformer):
_test_learning_kernel_with_random_feature('kl', transformer, rho=0.01)